Silver options: mock prices, model check
Package backend/silveropt/. It is not used by the Silver Arbitrage strategy any more: selling a strangle next to the short futures was
taken out of that backtest so it can become a separate strategy. What stays is reusable: the pricer, the volatility model and its calibration,
the mock EOD store, the real-price reader, the strike ladder and the strangle schedule builder (strangle.py). The engine side (premium and
payoff cash flows, option liability in equity, scan-risk margin, the page section) was removed; it is in git history, commit 09bc291, for the new
strategy to start from.
Why a model
MCX listed silver options in June 2018, but they were thin: SilverM option volume was 0.3 million units in 2022 and 2.3 million in 2023 (7.2 million in 2024). Real prices are therefore used from 2023-01-01; before that the backtest uses model prices, saved as end-of-day files in a store of their own. Real data is read from the real MCX F&O store and never copied into the mock store.
The model
- Black-76 on the Silver futures price (
black76.py), financing rate 6%, time to the option expiry. - Volatility (
volmodel.py): causal EWMA (lambda 0.94) volatility of the spot silver price, annualised, times a multipliermult(z)wherez = ln(K / F) / (sigma sqrt(T))is the strike distance in sigma units.mult(z)is the median of (implied volatility / realised volatility) per z bucket on real traded SilverM option prices (calls at or above the futures, puts at or below). Result: implied volatility trades at about 1.16 times the realised volatility on average, 1.05 near the money and 1.4-1.6 in the wings (a smile). - Cycle and strikes (
ladder.py): the option expires 16 days before its Silver futures contract (Mar / May / Jul / Sep / Dec cycle, as the real 2021-2025 SilverM options do), strikes on a ladder of about 0.3% of the futures price (250 at 110,000 like the real chain).
How good is it (out of sample)
Fitted on real prints from 2023-2024 (8,376 prints), tested on 2025-01-01 to 2026-10-23 (8,021 prints it never saw):
| median error | typical single price off by | within 50% | |
|---|---|---|---|
| plain realised volatility (no calibration) | -32.5% | 42% | 57% |
| calibrated model | +0.5% | 31% | 69% |
By strike distance (test window): at the money 21% typical error (87% within 50%), near 32%, far 47%, and the far wings (more than 3 sigma) are
underpriced by about 70%. By year: -10% in 2025, +14% in 2026 (the model lags volatility regime changes). Reproduce with
python -m silveropt.calibrate [--write]. The saved curve (config/calibration.json) is refitted on all 16,397 liquid real prints.
Reading it: the model is right on average and decent near the money, but any single early-year price is only roughly right. Pre-2023 results are an indication, not a record.
The mock store
cd backend
python -m silveropt.mock_store build --from 2006-01-01 --to 2022-12-31 # about 12 seconds
python -m silveropt.mock_store status
Config.MCXOPTMOCK_DATA_PATH (LOCAL_ / VPS_MCXOPTMOCK_DATA_PATH): raw/{yyyy}/{yyyymm} monthly Parquet, _catalog/manifest.json. EOD only,
columns sdate, expiry, fut_contract, strike, opttyp, fut_settle, dte, sigma, settle, delta. 1.78 million rows for 2006-2022 (strikes from
-20% to +20% of the futures price, every listed expiry within 100 days). The backtest recomputes the same model prices with the same code
(so a different strike distance or volatility multiplier works); a test checks that the stored rows equal the model.
The strangle schedule (strangle.py)
For every option cycle (entry on the first session after the previous expiry) strangle.build() picks an OTM call and/or put a given % away from the
futures price (legs both / call / put) and returns per-day arrays: entry premium, daily value, expiry payoff and a margin scan. It depends only on
market data, not on any account.
- Premium is credited to cash on entry; the open options' value is a liability in equity and in available margin; at expiry the intrinsic value against the futures settle is paid in cash (MCX exercises into futures; closing that futures the next day costs about the same, ignored).
- Margin is the worst loss of the whole book (short futures plus short options) after an up and a down move of the estimated margin percentage, so a short put offsets the short futures on a fall. An estimate, not exchange data.
- Costs: brokerage per leg, CTT 0.05% and exchange charge 0.0418% of the premium (reference values in
costs.yaml, applied to the mock years too; verify). - Real era (from 2023): the nearest listed real strike and the real settles are used for the entry premium and the daily value; a day without a print keeps the model path scaled to the last real print. 20 of 105 cycles are real, 85 model.
What it showed inside the Silver Arbitrage backtest (the earlier default plan 64 physical / 30 ETF / 6 cash / 30 short, 2006-2026, pledge credit 75%), for the new strategy to start from
| options sold | CAGR | max drawdown | hedged sleeve CAGR | option net |
|---|---|---|---|---|
| none | 13.6% | -28.6% | 3.2% | - |
| 5% strangle, 30 kg | 12.7% | -40.1% | -34.9% | +10.5 lakh, 59 kg of physical silver sold |
| 10% strangle, 10 kg | 13.7% | -27.6% | -8.1% | +4.1 lakh |
| 15% strangle, 30 kg | 14.2% | -27.3% | -13.8% | +11.0 lakh |
| put only, 10%, 30 kg | 14.2% | -32.8% | 4.2% | +20.8 lakh, 86% of cycles won |
| put only, 15%, 30 kg | 14.5% | -29.7% | 11.8% | +16.8 lakh, 95% won |
| call only, 10%, 30 kg | 13.0% | -25.6% | -23.5% | -8.6 lakh |
- The call leg loses money (silver trended up): every call-only row has negative net option income. Next to a short futures a call is short silver twice, so a 5% strangle on the full 30 kg cuts the CAGR and deepens the drawdown (-40%).
- The put leg earns because silver rose over the period; it would have lost in a long fall. Most of the income comes from the real-price years (2023-2026, premiums were very high in the 2025-26 rally); the 85 mock cycles before 2023 add little.
- Hedged sleeve: the sleeve pays every option payoff, so a strangle usually leaves the hedged-sleeve CAGR negative even when the whole strategy gains; only the put-only rows keep it above zero (the pass rule asks for >= 0).
- The earlier version of this table used a 25% pledge credit by mistake (see Silver Arbitrage > Pledge haircut) and showed the 5% strangle at -79%; with the broker's real collateral value it is -40%.
- These are results of one history with model prices for the early years; read them with the accuracy table above.